Initial Background Research
Finding Exoplanets Through Neural Networks
Neural Networks
Exoplanets
Planet outside the solar system
Tracked by Kepler Space Telescope
4,005 confirmed planets
1,077 TESS Project Candidates
Set of Algorithms designed to recognize patterns
Designed loosely after human brain.
Classification and Clustering
Perceptron
Collection of weights ordered in layers
Input layer
Receives series of numerical values
Place of weights(nodes)
Values multiplied by weights, go to next layer
Keeps going until output layer
Output is set of numerical values
Simplest type of neural network
Input values X = [x1, …]
N inputs, n weights
Weights = [w1]
Output si the dot product of two matrices
Z = XW = x1w1…
click to edit
The value z goes through activation function
Sigmoid function
Normalization of output
ReLU function
Final output A
Weights have to change for output to match the expected value
Neural Network has to trained
Back propagation
Labeled data
Numerical inputs where output is known
Initialize the weights, plug in random values
Labeled data goes through network, output compared to expected output
Find error through loss function
Binary cross entropy loss function(Keras library)
Adam optimizer
Dropout Layers
Deactivated random connections during training
Nodes more independent
First layer of CNN bust be convolutional
Generates feature maps, detect patterns
Use light curves
Light curves that represent known stars and exoplanets can be used as training data
Brightness levels of stars, dips because something(exoplanets) covers the spot between the star and observer
Mikulski Archive for Space Telescopes used to download data from Kepler Mission
Presearch Data Conditioning–Maximum A Posteriori (PDCMAP) used to calibrate, find apertures, aperture photometry, remove instrumental artifacts
Used as inputs for neural network
Lots of False Positives
click to edit
Similar radii to planets
Machine learning can classify exoplanets and false positives
Methods used are Random Forest, Dimensionality Reduction
Technique, K-Nearest Neighbours(KNN), Self Organising Maps and Neural Networks.
click to edit
Classification of Exoplanet Light Curves
Python
Oversampling Technique
SMOTE
Synthetic Minority Oversampling Technique
Imbalance in data
Normalization and Standardization
Get signals in uniform scale for comparison
Savitzky Golay Filter
Smoothing filter
Boots performance of classification technique
click to edit
Implementation
Data Pre-processing
Normalizing
Gaussian Filtering
Savitzky Golay Filter
Standardizing
Model Building
Input layer(ReLU Activation Function)
Hidden layer(ReLu activation function)
Output Classification Layer(Sigmoid activation function
Hyperparemters
Learning Rate(0.001)
Drop-out rate()
Epoch = 100
Batch size = 32
Optimizer
Stochastic Gradient Descent algorithm(SGD)
Find set of optimal weights
click to edit
Exoplanet Detection
Transit Method
Looks at the brightness of a star at regular intervals
Brightness vs time = light curve
Transiting planets had repeating dips in brightness
Planet blocks brightness between observer and star so dip in brightness
Size of dip = ratio between frontal surfaces of parent star and planet
click to edit
Probability of Habitability for Exoplanets
Classification into Mesoplanet, Psychroplanet, Non-habitable
Based on thermal temperatures of Exoplanets
Planetary Habitability Laboratory
Exoplanet catalog
Mesoplanets
Smaller than Mercury but larger than Ceres
Support life 10-45 degrees Celsius
Psychroplanets
-50-0 degrees celsius
Psychrophiles deep inside ice of Antarctica
Non-Habitable Planets
Don’t belong to either category
click to edit
Habitability of Exoplanets
Need water in liquid form
Forms by being appropriate distance away from parent star
Known as Habitable or Goldilocks Zone
0.95 AU to 1.97 AU
Astronomical Unit
Distance away from Earth to Sun
Atmospheric Composition is also very important
Venus and Mars are in in Habitable Zone but atmospheres don’t work
Size, Radius, Mass, Orbit also important
0.5-5 times as big as Earth
Radius is 0.8-1.5 times radius of earth